›› 2017, Vol. 60 ›› Issue (11): 1339-1348.doi: 10.16380/j.kcxb.2017.11.012

• 研究论文 • 上一篇    下一篇

基于深度学习的蝴蝶科级标本图像自动识别

周爱明1, 马鹏鹏1, 席天宇2, 王江宁2, 冯晋1, 邵泽中1, 陶玉磊1, 姚青1,*   

  1.  (1. 浙江理工大学信息学院, 杭州 310016; 2. 中国科学院动物研究所, 北京 100101)
  • 出版日期:2017-11-20 发布日期:2017-11-20

Automatic identification of butterfly specimen images at the family level based on deep learning method

ZHOU Ai-Ming1, MA Peng-Peng1, XI Tian-Yu2, WANG Jiang-Ning2, FENG Jin1, SHAO Ze-Zhong1, TAO Yu-Lei1, YAO Qing1,*   

  1. (1. School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China; 2. Institute of Zoology, Chinese Academy of Sciences, Beijing 100101, China)
  • Online:2017-11-20 Published:2017-11-20

摘要: 【目的】本研究旨在探讨深度学习模型在蝴蝶科级标本图像自动识别中的可行性和泛化能力。【方法】为了提高识别模型的鲁棒性和泛化能力,将锤角亚目中6个科1 117种蝴蝶标本图像通过水平翻转、增加图像对比度与亮度以及添加噪声的方式增强图像数据集。在Caffe框架下,利用迁移学习方法,首先使用ImageNet数据集中的图像训练CaffeNet模型,迭代31万次后得到初始化的网络权值;然后利用蝴蝶图像训练已预训练好的CaffeNet模型,通过参数微调,获得一个蝴蝶科级标本图像自动识别的卷积神经网络模型。为了比较深度学习和传统模式识别两种方法建立的模型的泛化能力,对相同训练样本提取全局特征和局部特征,训练支持向量机(support vector machine, SVM)分类器。所有的模型在与训练样本图像来源一致和不一致的两个测试样本集上进行测试。【结果】当测试样本与训练样本来源一致,均为蝴蝶标本图像时,基于CaffeNet的蝴蝶识别模型对6个科的蝴蝶识别准确率平均达到95.8%,基于Gabor的SVM分类器也获得了94.8%的识别率。当测试样本与训练样本来源不一致,为自然环境下拍摄的蝴蝶图像时,两种方法获得的识别率均下降,但CaffeNet模型对蝴蝶自然图像的平均识别率仍能达到65.6%,而基于Gabor的SVM分类器的识别率仅为38.9%。【结论】利用CaffeNet模型进行蝴蝶科级标本图像识别是可行的,相比较传统模式识别方法,基于深度学习的蝴蝶识别模型具有更好的泛化能力。

关键词: 蝴蝶, 标本图像, 自动识别, 深度学习, CaffeNet模型, 特征提取, 支持向量机

Abstract: 【Aim】 This study aims to explore the feasibility and generalization ability of deep learning model applied to the automatic identification of butterfly images at the family level. 【Methods】 To improve the robustness and generalization performance of model, the data augmentation with images of 1 117 butterfly species of six families were performed to increase the number of images by flipping image horizontally, increasing image contrast and brightness, and adding noises for training. In Caffe framework, an ImageNet-trained convolution neural network model was obtained by 310 000 iterations. The training set of butterfly images was used to train a new CaffeNet model to automatically identify butterflies at the family level by the transfer learning method. To compare generalization ability of the CaffeNet model based on deep learning with the models based on traditional pattern recognition methods, global and local features were extracted from the same training samples, and the support vector machine (SVM) classifier was trained. All models were used to detect the two different test sample sets. 【Results】 When the test samples, same as the training samples, were from specimen images, the CaffeNet model had a mean accuracy rate of 95.8%, while the SVM classifier based on Gabor features had a mean accuracy rate of 94.8% in six butterfly families. When the test samples were from natural images of butterflies, the accuracy rates of the CaffeNet and SVM models were decreased. However, the accuracy rate of CaffeNet model still achieved 65.6% and the SVM classifier based on Gabor features only got the 38.9% accuracy rate. 【Conclusion】 The butterfly identification model based on deep learning has a high identification rate at the family level, with higher robustness and generalization ability than those traditional pattern recognition models based on global and local features by manual extraction and selection.

Key words: Butterfly, specimen images, automatic identification, deep learning, CaffeNet model, feature extraction, support vector machine